Often we have to handle high dimensional spaces if we want to learn motor skills for robots. In policy search tasks we have to find several parameters to learn a desired movement. This high dimensionality in parameters can be challenging for reinforcement algorithms, since more samples for finding an optimal solution are needed with every additional dimension. On the other hand, if the robot has a high number of actuators, an inherent correlation between these can be found for a specific motor task, which we can exploit for a faster convergence. One possibility is to use techniques to reduce the dimensionality of the space, which is used as a pre-processing step or as an independent process in most applications. In this thesis we present a novel algorithm which combines the theory of policy search and probabilistic dimensionality reduction to uncover the hidden structure of high dimensional action spaces. Evaluations on an inverse kinematics task indicate that the presented algorithm is able to outperform the reference algorithms PoWER and CMA-ES, especially in high dimensional spaces. Furthermore we evaluate our algorithm on a real-world task. In this task, a NAO robot learns to lift his leg while keeping balance. The issue of collecting samples for learning on a real robot in such a task, which is often very time and cost consuming, is considered in here by using a small number of samples in each iteration.
Latent Space Reinforcement Learning
2014-05-06
Theses
Electronic Resource
German , English
DDC: | 629 |
Do Androids Dream of Electric Fences? Safety-Aware Reinforcement Learning with Latent Shielding
BASE | 2022
|Rotterdam Ausstellung. Latent Space
Online Contents | 2002